microsoft / deep-language-networks

We view Large Language Models as stochastic language layers in a network, where the learnable parameters are the natural language prompts at each layer. We stack two such layers, feeding the output of one layer to the next. We call the stacked architecture a Deep Language Network - DLN
MIT License
91 stars 13 forks source link

Remove weird tokenization artifacts from vLLM models #45

Closed matheper closed 8 months ago